gin: add GDAKI plugin with stub implementations

include/Makefile.am

@@ -14,6 +14,7 @@ noinst_HEADERS = \
 	rdma/gin/nccl_ofi_gin_reqs.h \
 	rdma/gin/nccl_ofi_gin_resources.h \
 	rdma/gin/nccl_ofi_gin_types.h \
+	rdma/gin/nccl_ofi_gin_gdaki.h \
 	nccl_ofi_gin_base.h \
 	nccl_ofi.h \
 	nccl_ofi_api.h \

include/nccl_ofi_param.h

@@ -374,4 +374,10 @@ OFI_NCCL_PARAM(std::string, platform, "PLATFORM", "");
 OFI_NCCL_PARAM_VALUE_SET(NVTX_TRACE_DIMENSION, (PER_COMM)(PER_DEV))
 OFI_NCCL_PARAM(NVTX_TRACE_DIMENSION, nvtx_trace_dimension,  "NVTX_TRACE_DIMENSION", NVTX_TRACE_DIMENSION::PER_COMM)
 
+/*
+ * Enable GDAKI (GPUDirect Async) mode for the GIN plugin.
+ * Not yet supported — stub implementation only.
+ */
+OFI_NCCL_PARAM(bool, gin_gdaki, "GIN_GDAKI", false);
+
 #endif // End NCCL_OFI_PARAM_H_

include/rdma/gin/nccl_ofi_gin_gdaki.h

@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2026 Amazon.com, Inc. or its affiliates. All rights reserved.
+ */
+
+#ifndef NCCL_OFI_GIN_GDAKI_H_
+#define NCCL_OFI_GIN_GDAKI_H_
+
+#include "nccl_ofi.h"
+
+/*
+ * Return true if GDAKI mode is requested via OFI_NCCL_GIN_GDAKI=1 env var.
+ */
+bool nccl_ofi_gin_gdaki_enabled();
+
+/*
+ * The GDAKI plugin. Shared functions (init, devices, listen, connect)
+ * are nullptr and get copied from the proxy plugin at init time.
+ */
+extern ncclGin_v11_t nccl_ofi_gin_gdaki_plugin;
+
+#endif /* NCCL_OFI_GIN_GDAKI_H_ */

src/Makefile.am

@@ -56,7 +56,8 @@ sources += \
 	rdma/gin/nccl_ofi_gin_allgather.cpp \
 	rdma/gin/nccl_ofi_gin_api.cpp \
 	rdma/gin/nccl_ofi_gin_reqs.cpp \
-	rdma/gin/nccl_ofi_gin_resources.cpp
+	rdma/gin/nccl_ofi_gin_resources.cpp \
+	rdma/gin/nccl_ofi_gin_gdaki.cpp
 endif
 
 # add the tuner sources into the library

src/rdma/gin/nccl_ofi_gin_api.cpp

@@ -4,12 +4,18 @@
 
 #include "config.h"
 
+#include <cstring>
+
 #include "rdma/gin/nccl_ofi_gin.h"
 #include "rdma/gin/nccl_ofi_gin_types.h"
+#include "rdma/gin/nccl_ofi_gin_gdaki.h"
 #include "nccl_ofi.h"
 #include "nccl_ofi_api.h"
 #include "nccl_ofi_param.h"
 
+/* Forward declaration — defined at bottom of this file */
+extern ncclGin_v11_t ncclGinPlugin_v11;
+
 /**
  * Structure to hold GIN context data.
  * This is created once per NCCL communicator and passed to all listen() calls
@@ -69,6 +75,22 @@ static ncclResult_t nccl_ofi_gin_init(void **ctx, uint64_t commId, ncclDebugLogg
 		return ncclSystemError;
 	}
 
+	/*
+	 * Morph the exported plugin to GDAKI if requested.
+	 *
+	 * Copy shared functions (init, devices, listen, connect) from the
+	 * proxy plugin into the GDAKI plugin, then overwrite the exported
+	 * symbol with the GDAKI plugin.
+	 */
+	if (nccl_ofi_gin_gdaki_enabled()) {
+		NCCL_OFI_INFO(NCCL_NET | NCCL_INIT, "gin: GDAKI mode enabled (OFI_NCCL_GIN_GDAKI=1)");
+		nccl_ofi_gin_gdaki_plugin.init = ncclGinPlugin_v11.init;
+		nccl_ofi_gin_gdaki_plugin.devices = ncclGinPlugin_v11.devices;
+		nccl_ofi_gin_gdaki_plugin.listen = ncclGinPlugin_v11.listen;
+		nccl_ofi_gin_gdaki_plugin.connect = ncclGinPlugin_v11.connect;
+		memcpy(&ncclGinPlugin_v11, &nccl_ofi_gin_gdaki_plugin, sizeof(ncclGinPlugin_v11));
+	}
+
 	return ncclSuccess;
 }
 

src/rdma/gin/nccl_ofi_gin_gdaki.cpp

@@ -0,0 +1,151 @@
+/*
+ * Copyright (c) 2026 Amazon.com, Inc. or its affiliates. All rights reserved.
+ *
+ * GDAKI stub implementations for the GIN plugin API.
+ * These provide the full ncclGin_v11_t plugin for GDAKI mode.
+ *
+ * Task: GDAKI stub implementation
+ */
+
+#include "config.h"
+
+#include "rdma/gin/nccl_ofi_gin_gdaki.h"
+#include "nccl_ofi.h"
+#include "nccl_ofi_api.h"
+#include "nccl_ofi_param.h"
+
+bool nccl_ofi_gin_gdaki_enabled()
+{
+	return ofi_nccl_gin_gdaki.get();
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_get_properties(int dev, ncclNetProperties_v11_t *props)
+{
+	nccl_ofi_properties_t ofi_properties;
+	ncclResult_t ret = nccl_net_ofi_get_properties(dev, &ofi_properties);
+	if (ret != ncclSuccess) {
+		return ret;
+	}
+
+	props->name = ofi_properties.name;
+	props->pciPath = ofi_properties.pci_path;
+	props->guid = ofi_properties.guid;
+	props->ptrSupport = NCCL_PTR_HOST;
+	if (ofi_properties.hmem_support) {
+		props->ptrSupport |= NCCL_PTR_CUDA;
+	}
+	if (ofi_properties.dmabuf_support) {
+		props->ptrSupport |= NCCL_PTR_DMABUF;
+	}
+
+	props->regIsGlobal = ofi_properties.regIsGlobal;
+	props->speed = ofi_properties.port_speed;
+	props->port = ofi_properties.port_number;
+	props->latency = ofi_properties.latency;
+	props->maxComms = ofi_properties.max_communicators;
+	props->maxRecvs = ofi_properties.max_group_receives;
+	props->netDeviceType = NCCL_NET_DEVICE_GIN_GDAKI;
+	props->netDeviceVersion = NCCL_NET_DEVICE_INVALID_VERSION;
+	props->vProps.ndevs = 1;
+	props->vProps.devs[0] = dev;
+	props->maxP2pBytes = ofi_properties.max_p2p_bytes;
+	props->maxCollBytes = ofi_properties.max_coll_bytes;
+
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_createContext(void *collComm, int nSignals, int nCounters,
+						     void **ginCtx,
+						     ncclNetDeviceHandle_v11_t **devHandle)
+{
+	NCCL_OFI_WARN("gin GDAKI: createContext not yet implemented (nSignals=%d, nCounters=%d)",
+		      nSignals, nCounters);
+	return ncclInternalError;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_destroyContext(void *ginCtx)
+{
+	NCCL_OFI_WARN("gin GDAKI: destroyContext not yet implemented");
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_regMrSym(void *collComm, void *data, size_t size, int type,
+						 uint64_t mrFlags, void **mhandle, void **ginHandle)
+{
+	NCCL_OFI_WARN("gin GDAKI: regMrSym not yet implemented");
+	return ncclInternalError;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_regMrSymDmaBuf(void *collComm, void *data, size_t size,
+						       int type, uint64_t offset, int fd,
+						       uint64_t mrFlags, void **mhandle,
+						       void **ginHandle)
+{
+	NCCL_OFI_WARN("gin GDAKI: regMrSymDmaBuf not yet implemented");
+	return ncclInternalError;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_deregMrSym(void *collComm, void *mhandle)
+{
+	NCCL_OFI_WARN("gin GDAKI: deregMrSym not yet implemented");
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_closeColl(void *collComm)
+{
+	NCCL_OFI_WARN("gin GDAKI: closeColl not yet implemented");
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_closeListen(void *listenComm)
+{
+	NCCL_OFI_WARN("gin GDAKI: closeListen not yet implemented");
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_ginProgress(void *collComm)
+{
+	NCCL_OFI_WARN("gin GDAKI: ginProgress not yet implemented");
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_queryLastError(void *ginCtx, bool *hasError)
+{
+	NCCL_OFI_WARN("gin GDAKI: queryLastError not yet implemented");
+	*hasError = false;
+	return ncclSuccess;
+}
+
+static ncclResult_t nccl_ofi_gin_gdaki_finalize(void *ctx)
+{
+	NCCL_OFI_WARN("gin GDAKI: finalize not yet implemented");
+	return ncclSuccess;
+}
+
+/*
+ * GDAKI plugin. Function pointers for the ncclGin_v11_t interface:
+ * - iput, iputSignal, test are nullptr (no CPU involvement in GDAKI mode)
+ * - init, devices, listen, connect are copied from the proxy plugin at
+ *   init time.
+ */
+ncclGin_v11_t nccl_ofi_gin_gdaki_plugin = {
+	.name = "Libfabric_GDAKI",
+	.init = nullptr,	/* Copied from proxy plugin at init time */
+	.devices = nullptr,	/* Copied from proxy plugin at init time */
+	.getProperties = nccl_ofi_gin_gdaki_get_properties,
+	.listen = nullptr,	/* Copied from proxy plugin at init time */
+	.connect = nullptr,	/* Copied from proxy plugin at init time */
+	.createContext = nccl_ofi_gin_gdaki_createContext,
+	.regMrSym = nccl_ofi_gin_gdaki_regMrSym,
+	.regMrSymDmaBuf = nccl_ofi_gin_gdaki_regMrSymDmaBuf,
+	.deregMrSym = nccl_ofi_gin_gdaki_deregMrSym,
+	.destroyContext = nccl_ofi_gin_gdaki_destroyContext,
+	.closeColl = nccl_ofi_gin_gdaki_closeColl,
+	.closeListen = nccl_ofi_gin_gdaki_closeListen,
+	.iput = nullptr,	/* No CPU involvement in GDAKI mode */
+	.iputSignal = nullptr,	/* No CPU involvement in GDAKI mode */
+	.test = nullptr,	/* No CPU involvement in GDAKI mode */
+	.ginProgress = nccl_ofi_gin_gdaki_ginProgress,
+	.queryLastError = nccl_ofi_gin_gdaki_queryLastError,
+	.finalize = nccl_ofi_gin_gdaki_finalize
+};